Advanced YOLOv4 for real-time underwater object detection: An application-oriented approach

Detecting debris and monitoring marine life in sea aquaculture face challenges due to limited visibility and the presence of diverse. Underwater object detection by Autonomous Unmanned Vehicle(AUV) is inherently more challenging than land due to light attenuation and water turbidity, especially for small and dense objects in murky images, where extracting high-quality features is hindered. In this paper, we present an efficient approach for real-time underwater object detection through improvements in image enhancement, data augmentation, and feature aggregation. Initially, U-Shape Transformer is applied to enhance the original images. For data augmentation, it is observable that while Mosaic data augmentation enhances complex images but fails to improve small-object detection due generation of less number of images with small objects. To address this limitation, we propose Underwater-Mosaic (U-Mosaic), a modified Mosaic data augmentation technique designed to enhance small-object detection. Additionally, it was noted that existing YOLOv4 struggles with detecting small and densely populated objects in underwater images as unable to get sufficient features for small objects due to downsampling, image quality and also found difficulty in selecting anchor box size. Therefore, we propose a model called Advanced YOLOv4, tailored for underwater object detection. The proposed Advanced YOLOv4 aims to improve object detection efficiency by altering the neck and prediction layers of YOLOv4. Moreover, we introduce an additional spatial pyramid pooling layer to aggregate features and reduce feature dimensions thereby improving object detection rates. Also, the proposed work concentrates on very large object detection and for this purpose used downsampling during the detection of large objects. The proposed approach is validated through two distinct application areas: (i) detecting and locating debris (ii) detecting fish from underwater images. For validation, the Trash ICRA19 dataset is used for debris detection, while the Brackish dataset is employed for fish detection. UIQM and UCIQE, image enhancement assessment metrics are used to measure quality of enhanced images and found more than 20% better result for both the datasets. The proposed real-time underwater object detection model outperformed single-stage object detectors like YOLOv3, YOLOv4, YOLOv5, YOLOv7, and KPE-YOLOv5 by 5% in terms of mean Average Precision(mAP). Also proposed work compared with two-stage detector RCNN and found 8% better mAP than RCNN.